Systems and methods for accelerated magnetic resonance imaging (MRI) reconstruction and sampling

Abstract

The following relates generally to accelerated magnetic resonance imaging (MRI) reconstruction. In some embodiments, a MRI machine learning algorithm is trained based on reference MRI data in non-Cartesian k-space. During the training, at each iteration of a plurality of iterations: (i) a non-Cartesian sampling trajectory ω may be optimized under the physical constraints, and/or (ii) an image reconstructor may be jointly iteratively optimized. Examples of the image reconstructor include a convolutional neural network (CNN) denoiser, a model-based deep learning (MoDL) image reconstructor, iterative image reconstructor, a regularizer, and an invertible neural network.

What is claimed: 
     
       1. A computer-implemented method for designing a non-Cartesian sampling trajectory for either a prespecified image reconstructor or an optimized image reconstructor for producing a magnetic resonance imaging (MRI) image, the method comprising:
 training, via one or more processors, a MRI machine learning model to design a non-Cartesian MRI sampling trajectory for either the prespecified image reconstructor or for the optimized image reconstructor for producing an MRI image; 
 parameterizing, by the one or more processors, the non-Cartesian sampling trajectory using a basis function set; 
 generating, by the one or more processors, the non-Cartesian sampling trajectory for imaging a patient using the MRI machine learning model; 
 generating, by the one or more processors, MRI data for the patient using the non-Cartesian sampling trajectory; 
 reconstructing, by the one or more processors, the MRI data using either prespecified reconstructor or the optimized image reconstructor; and 
 storing, by the one or more processors, the reconstructed MRI data in a memory. 
 
     
     
       2. A device for designing a non-Cartesian sampling trajectory for either a prespecified image reconstructor or an optimized image reconstructor for producing a magnetic resonance imaging (MRI) image, the device comprising one or more processors configured to:
 train a MRI machine learning model to design a non-Cartesian MRI sampling trajectory for either the prespecified image reconstructor or for the optimized image reconstructor for producing an MRI image; 
 parameterize the non-Cartesian sampling trajectory using a basis function set; 
 generate the non-Cartesian sampling trajectory for imaging a patient using the MRI machine learning model; 
 generate MRI data for the patient using the non-Cartesian sampling trajectory; 
 reconstruct the MRI data using either prespecified reconstruction parameters or optimized image reconstruction parameters; and 
 store the reconstructed MRI data in a database. 
 
     
     
       3. A system for designing a non-Cartesian sampling trajectory for either a prespecified image reconstructor or an optimized image reconstructor for producing a magnetic resonance imaging (MRI) image, the system comprising:
 one or more processors; 
 at least one transmit coil; 
 at least one receive coil; 
 a user interface; and 
 one or more memories coupled to the one or more processors; 
 the one or more memories including computer executable instructions stored therein that, when executed by the one or more processors, cause the one or more processors to:
 train a MRI machine learning model to design a non-Cartesian MRI sampling trajectory for either the prespecified image reconstructor or for the optimized image reconstructor for producing an MRI image; 
 parameterize the non-Cartesian sampling trajectory using a basis function set; 
 generate the non-Cartesian sampling trajectory for imaging a patient using the MRI machine learning model; 
 generate MRI data for the patient using the non-Cartesian sampling trajectory; 
 reconstruct the MRI data using either prespecified reconstruction parameters or reconstruction parameters to produce the MRI image; and 
 display the MRI image on the user interface. 
 
 
     
     
       4. The computer-implemented method of  claim 1 , wherein training the MRI machine learning model includes:
 obtaining, by the one or more processors, a training set of MRI data for an MRI image; 
 selecting, by the one or more processors, a non-Cartesian sampling trajectory (ω) for the training set; 
 initializing, by the one or more processors, a reconstruction of the training set of MRI data using an initial set of reconstruction parameters; and 
 iteratively updating the reconstruction by updating a reconstruction parameter (θ) of a convolutional neural network (CNN) denoiser and updating the non-Cartesian sampling trajectory (ω) to minimize a difference between a ground truth MRI image and the reconstruction. 
 
     
     
       5. The computer-implemented method of  claim 1 , wherein training the MRI machine learning model to design the non-Cartesian sampling trajectory comprises penalizing the non-Cartesian sampling trajectory to obey hardware constraints, the hardware constraints comprising maximum slew rate and gradient strength. 
     
     
       6. The computer-implemented method of  claim 1 , wherein the basis function set comprises second-order quadratic B-spline kernels. 
     
     
       7. The device of  claim 2 , wherein the non-Cartesian sampling trajectory is step-wise differentiable, thereby enabling differentiable programming. 
     
     
       8. The device of  claim 2 , wherein the one or more processors are configured to train the MRI machine learning model according to:
     =∥ f   θ,ω ( A (ω) x +ε)− x∥ 
 
 where: 
    is loss during reconstruction during the training; 
 ∥·∥ is a norm comprising a loss function that compares a reconstructed image to a training image; 
 θ is the reconstruction parameters; 
 ω is the non-Cartesian sampling trajectory; 
 A(ω) denotes a system matrix for the non-Cartesian sampling trajectory ω; 
 x is training data; and 
 ε is simulated additive noise. 
 
     
     
       9. The device of  claim 2 , wherein the image reconstructor is part of an unrolled neural network, and the unrolled neural network comprises the image reconstructor, and a data consistency unit. 
     
     
       10. The device of  claim 2 , the one or more processors further configured to train the MRI machine learning model by:
 obtaining a training set of MRI data for an MRI image; 
 using a data consistency unit to compare data produced by the image reconstructor to the training set of MRI data; and 
 if the data produced by the image reconstructor and the training set of MRI data differ by more than a predetermined amount, updating the reconstruction parameters. 
 
     
     
       11. The device of  claim 2 , wherein the one or more processors are further configured to train by optimizing trajectory attributes (c), wherein ω is the non-Cartesian sampling trajectory, and ω(c) is a nonlinear function of the trajectory attributes (c). 
     
     
       12. The system of  claim 3 , wherein the one or more memories including computer executable instructions stored therein that, when executed by the one or more processors, further cause the one or more processors to:
 generate the MRI data by controlling the transmit coil and the receive coil to acquire the MRI data according to the non-Cartesian sampling trajectory. 
 
     
     
       13. The system of  claim 3 , wherein:
 the image reconstructor is an invertible neural network; and 
 the one or more memories including computer executable instructions stored therein that, when executed by the one or more processors, further cause the one or more processors to generate the MRI data for the patient in non-Cartesian three-dimensional (3D) k-space. 
 
     
     
       14. The system of  claim 3 , wherein:
 the one or more memories including computer executable instructions stored therein that, when executed by the one or more processors, further cause the one or more processors to generate the MRI data for the patient in non-Cartesian k-space, the non-Cartesian k-space having a plurality of dimensions; and 
 no dimension of the plurality of dimensions has a least common divisor. 
 
     
     
       15. The system of  claim 3 , wherein:
 the image reconstructor is a regularizer; and 
 the one or more memories including computer executable instructions stored therein that, when executed by the one or more processors, further cause the one or more processors to train the MRI machine learning model by updating a proximal operator of the regularizer. 
 
     
     
       16. The system of  claim 3 , wherein the image reconstructor is a model-based deep learning (MoDL) image reconstructor. 
     
     
       17. The computer-implemented method of  claim 4 , wherein the MRI data is generated using the updated non-Cartesian sampling trajectory (ω) and the MRI data is reconstructed using the updated reconstruction parameter (θ). 
     
     
       18. The computer-implemented method of  claim 4 , wherein iteratively updating the reconstruction includes increasing a number of basis functions of the basis function set, beginning a new round of training, and refining the reconstruction parameter (θ). 
     
     
       19. The computer-implemented method of  claim 4 , wherein a matrix for the training set is constructed using a non-uniform fast Fourier transform (NUFFT). 
     
     
       20. The computer-implemented method of  claim 4 , wherein the reconstruction parameter (θ) and the non-Cartesian sampling trajectory (ω) are designed simultaneously.